1/* 2 * Copyright (c) 2013 ARM Ltd 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. The name of the company may not be used to endorse or promote 14 * products derived from this software without specific prior written 15 * permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED 18 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED 22 * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 23 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 24 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 25 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 26 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29#include <machine/cpu-features.h> 30#include <private/bionic_asm.h> 31 32#ifdef __ARMEB__ 33#define S2LOMEM lsl 34#define S2LOMEMEQ lsleq 35#define S2HIMEM lsr 36#define MSB 0x000000ff 37#define LSB 0xff000000 38#define BYTE0_OFFSET 24 39#define BYTE1_OFFSET 16 40#define BYTE2_OFFSET 8 41#define BYTE3_OFFSET 0 42#else /* not __ARMEB__ */ 43#define S2LOMEM lsr 44#define S2LOMEMEQ lsreq 45#define S2HIMEM lsl 46#define BYTE0_OFFSET 0 47#define BYTE1_OFFSET 8 48#define BYTE2_OFFSET 16 49#define BYTE3_OFFSET 24 50#define MSB 0xff000000 51#define LSB 0x000000ff 52#endif /* not __ARMEB__ */ 53 54.syntax unified 55 56#if defined (__thumb__) 57 .thumb 58 .thumb_func 59#endif 60 61ENTRY(strcmp) 62 /* Use LDRD whenever possible. */ 63 64/* The main thing to look out for when comparing large blocks is that 65 the loads do not cross a page boundary when loading past the index 66 of the byte with the first difference or the first string-terminator. 67 68 For example, if the strings are identical and the string-terminator 69 is at index k, byte by byte comparison will not load beyond address 70 s1+k and s2+k; word by word comparison may load up to 3 bytes beyond 71 k; double word - up to 7 bytes. If the load of these bytes crosses 72 a page boundary, it might cause a memory fault (if the page is not mapped) 73 that would not have happened in byte by byte comparison. 74 75 If an address is (double) word aligned, then a load of a (double) word 76 from that address will not cross a page boundary. 77 Therefore, the algorithm below considers word and double-word alignment 78 of strings separately. */ 79 80/* High-level description of the algorithm. 81 82 * The fast path: if both strings are double-word aligned, 83 use LDRD to load two words from each string in every loop iteration. 84 * If the strings have the same offset from a word boundary, 85 use LDRB to load and compare byte by byte until 86 the first string is aligned to a word boundary (at most 3 bytes). 87 This is optimized for quick return on short unaligned strings. 88 * If the strings have the same offset from a double-word boundary, 89 use LDRD to load two words from each string in every loop iteration, as in the fast path. 90 * If the strings do not have the same offset from a double-word boundary, 91 load a word from the second string before the loop to initialize the queue. 92 Use LDRD to load two words from every string in every loop iteration. 93 Inside the loop, load the second word from the second string only after comparing 94 the first word, using the queued value, to guarantee safety across page boundaries. 95 * If the strings do not have the same offset from a word boundary, 96 use LDR and a shift queue. Order of loads and comparisons matters, 97 similarly to the previous case. 98 99 * Use UADD8 and SEL to compare words, and use REV and CLZ to compute the return value. 100 * The only difference between ARM and Thumb modes is the use of CBZ instruction. 101 * The only difference between big and little endian is the use of REV in little endian 102 to compute the return value, instead of MOV. 103*/ 104 105 .macro m_cbz reg label 106#ifdef __thumb2__ 107 cbz \reg, \label 108#else /* not defined __thumb2__ */ 109 cmp \reg, #0 110 beq \label 111#endif /* not defined __thumb2__ */ 112 .endm /* m_cbz */ 113 114 .macro m_cbnz reg label 115#ifdef __thumb2__ 116 cbnz \reg, \label 117#else /* not defined __thumb2__ */ 118 cmp \reg, #0 119 bne \label 120#endif /* not defined __thumb2__ */ 121 .endm /* m_cbnz */ 122 123 .macro init 124 /* Macro to save temporary registers and prepare magic values. */ 125 subs sp, sp, #16 126 .cfi_def_cfa_offset 16 127 strd r4, r5, [sp, #8] 128 .cfi_rel_offset r4, 0 129 .cfi_rel_offset r5, 4 130 strd r6, r7, [sp] 131 .cfi_rel_offset r6, 8 132 .cfi_rel_offset r7, 12 133 mvn r6, #0 /* all F */ 134 mov r7, #0 /* all 0 */ 135 .endm /* init */ 136 137 .macro magic_compare_and_branch w1 w2 label 138 /* Macro to compare registers w1 and w2 and conditionally branch to label. */ 139 cmp \w1, \w2 /* Are w1 and w2 the same? */ 140 magic_find_zero_bytes \w1 141 it eq 142 cmpeq ip, #0 /* Is there a zero byte in w1? */ 143 bne \label 144 .endm /* magic_compare_and_branch */ 145 146 .macro magic_find_zero_bytes w1 147 /* Macro to find all-zero bytes in w1, result is in ip. */ 148 uadd8 ip, \w1, r6 149 sel ip, r7, r6 150 .endm /* magic_find_zero_bytes */ 151 152 .macro setup_return w1 w2 153#ifdef __ARMEB__ 154 mov r1, \w1 155 mov r2, \w2 156#else /* not __ARMEB__ */ 157 rev r1, \w1 158 rev r2, \w2 159#endif /* not __ARMEB__ */ 160 .endm /* setup_return */ 161 162 pld [r0, #0] 163 pld [r1, #0] 164 165 /* Are both strings double-word aligned? */ 166 orr ip, r0, r1 167 tst ip, #7 168 bne .L_do_align 169 170 /* Fast path. */ 171 .save {r4-r7} 172 init 173 174.L_doubleword_aligned: 175 176 /* Get here when the strings to compare are double-word aligned. */ 177 /* Compare two words in every iteration. */ 178 .p2align 2 1792: 180 pld [r0, #16] 181 pld [r1, #16] 182 183 /* Load the next double-word from each string. */ 184 ldrd r2, r3, [r0], #8 185 ldrd r4, r5, [r1], #8 186 187 magic_compare_and_branch w1=r2, w2=r4, label=.L_return_24 188 magic_compare_and_branch w1=r3, w2=r5, label=.L_return_35 189 b 2b 190 191.L_do_align: 192 /* Is the first string word-aligned? */ 193 ands ip, r0, #3 194 beq .L_word_aligned_r0 195 196 /* Fast compare byte by byte until the first string is word-aligned. */ 197 /* The offset of r0 from a word boundary is in ip. Thus, the number of bytes 198 to read until the next word boundary is 4-ip. */ 199 bic r0, r0, #3 200 ldr r2, [r0], #4 201 lsls ip, ip, #31 202 beq .L_byte2 203 bcs .L_byte3 204 205.L_byte1: 206 ldrb ip, [r1], #1 207 uxtb r3, r2, ror #BYTE1_OFFSET 208 subs ip, r3, ip 209 bne .L_fast_return 210 m_cbz reg=r3, label=.L_fast_return 211 212.L_byte2: 213 ldrb ip, [r1], #1 214 uxtb r3, r2, ror #BYTE2_OFFSET 215 subs ip, r3, ip 216 bne .L_fast_return 217 m_cbz reg=r3, label=.L_fast_return 218 219.L_byte3: 220 ldrb ip, [r1], #1 221 uxtb r3, r2, ror #BYTE3_OFFSET 222 subs ip, r3, ip 223 bne .L_fast_return 224 m_cbnz reg=r3, label=.L_word_aligned_r0 225 226.L_fast_return: 227 mov r0, ip 228 bx lr 229 230.L_word_aligned_r0: 231 init 232 /* The first string is word-aligned. */ 233 /* Is the second string word-aligned? */ 234 ands ip, r1, #3 235 bne .L_strcmp_unaligned 236 237.L_word_aligned: 238 /* The strings are word-aligned. */ 239 /* Is the first string double-word aligned? */ 240 tst r0, #4 241 beq .L_doubleword_aligned_r0 242 243 /* If r0 is not double-word aligned yet, align it by loading 244 and comparing the next word from each string. */ 245 ldr r2, [r0], #4 246 ldr r4, [r1], #4 247 magic_compare_and_branch w1=r2 w2=r4 label=.L_return_24 248 249.L_doubleword_aligned_r0: 250 /* Get here when r0 is double-word aligned. */ 251 /* Is r1 doubleword_aligned? */ 252 tst r1, #4 253 beq .L_doubleword_aligned 254 255 /* Get here when the strings to compare are word-aligned, 256 r0 is double-word aligned, but r1 is not double-word aligned. */ 257 258 /* Initialize the queue. */ 259 ldr r5, [r1], #4 260 261 /* Compare two words in every iteration. */ 262 .p2align 2 2633: 264 pld [r0, #16] 265 pld [r1, #16] 266 267 /* Load the next double-word from each string and compare. */ 268 ldrd r2, r3, [r0], #8 269 magic_compare_and_branch w1=r2 w2=r5 label=.L_return_25 270 ldrd r4, r5, [r1], #8 271 magic_compare_and_branch w1=r3 w2=r4 label=.L_return_34 272 b 3b 273 274 .macro miscmp_word offsetlo offsethi 275 /* Macro to compare misaligned strings. */ 276 /* r0, r1 are word-aligned, and at least one of the strings 277 is not double-word aligned. */ 278 /* Compare one word in every loop iteration. */ 279 /* OFFSETLO is the original bit-offset of r1 from a word-boundary, 280 OFFSETHI is 32 - OFFSETLO (i.e., offset from the next word). */ 281 282 /* Initialize the shift queue. */ 283 ldr r5, [r1], #4 284 285 /* Compare one word from each string in every loop iteration. */ 286 .p2align 2 2877: 288 ldr r3, [r0], #4 289 S2LOMEM r5, r5, #\offsetlo 290 magic_find_zero_bytes w1=r3 291 cmp r7, ip, S2HIMEM #\offsetlo 292 and r2, r3, r6, S2LOMEM #\offsetlo 293 it eq 294 cmpeq r2, r5 295 bne .L_return_25 296 ldr r5, [r1], #4 297 cmp ip, #0 298 eor r3, r2, r3 299 S2HIMEM r2, r5, #\offsethi 300 it eq 301 cmpeq r3, r2 302 bne .L_return_32 303 b 7b 304 .endm /* miscmp_word */ 305 306.L_return_32: 307 setup_return w1=r3, w2=r2 308 b .L_do_return 309.L_return_34: 310 setup_return w1=r3, w2=r4 311 b .L_do_return 312.L_return_25: 313 setup_return w1=r2, w2=r5 314 b .L_do_return 315.L_return_35: 316 setup_return w1=r3, w2=r5 317 b .L_do_return 318.L_return_24: 319 setup_return w1=r2, w2=r4 320 321.L_do_return: 322 323#ifdef __ARMEB__ 324 mov r0, ip 325#else /* not __ARMEB__ */ 326 rev r0, ip 327#endif /* not __ARMEB__ */ 328 329 /* Restore temporaries early, before computing the return value. */ 330 ldrd r6, r7, [sp] 331 ldrd r4, r5, [sp, #8] 332 adds sp, sp, #16 333 .cfi_def_cfa_offset 0 334 .cfi_restore r4 335 .cfi_restore r5 336 .cfi_restore r6 337 .cfi_restore r7 338 339 /* There is a zero or a different byte between r1 and r2. */ 340 /* r0 contains a mask of all-zero bytes in r1. */ 341 /* Using r0 and not ip here because cbz requires low register. */ 342 m_cbz reg=r0, label=.L_compute_return_value 343 clz r0, r0 344 /* r0 contains the number of bits on the left of the first all-zero byte in r1. */ 345 rsb r0, r0, #24 346 /* Here, r0 contains the number of bits on the right of the first all-zero byte in r1. */ 347 lsr r1, r1, r0 348 lsr r2, r2, r0 349 350.L_compute_return_value: 351 movs r0, #1 352 cmp r1, r2 353 /* The return value is computed as follows. 354 If r1>r2 then (C==1 and Z==0) and LS doesn't hold and r0 is #1 at return. 355 If r1<r2 then (C==0 and Z==0) and we execute SBC with carry_in=0, 356 which means r0:=r0-r0-1 and r0 is #-1 at return. 357 If r1=r2 then (C==1 and Z==1) and we execute SBC with carry_in=1, 358 which means r0:=r0-r0 and r0 is #0 at return. 359 (C==0 and Z==1) cannot happen because the carry bit is "not borrow". */ 360 it ls 361 sbcls r0, r0, r0 362 bx lr 363 364 /* The code from the previous version of strcmp.S handles all of the 365 * cases where the first string and seconds string cannot both be 366 * aligned to a word boundary faster than the new algorithm. See 367 * bionic/libc/arch-arm/cortex-a15/bionic/strcmp.S for the unedited 368 * version of the code. 369 */ 370.L_strcmp_unaligned: 371 wp1 .req r0 372 wp2 .req r1 373 b1 .req r2 374 w1 .req r4 375 w2 .req r5 376 t1 .req ip 377 @ r3 is scratch 378 3792: 380 mov b1, #1 381 orr b1, b1, b1, lsl #8 382 orr b1, b1, b1, lsl #16 383 384 and t1, wp2, #3 385 bic wp2, wp2, #3 386 ldr w1, [wp1], #4 387 ldr w2, [wp2], #4 388 cmp t1, #2 389 beq 2f 390 bhi 3f 391 392 /* Critical inner Loop: Block with 3 bytes initial overlap */ 393 .p2align 2 3941: 395 bic t1, w1, #MSB 396 cmp t1, w2, S2LOMEM #8 397 sub r3, w1, b1 398 bic r3, r3, w1 399 bne 4f 400 ands r3, r3, b1, lsl #7 401 it eq 402 ldreq w2, [wp2], #4 403 bne 5f 404 eor t1, t1, w1 405 cmp t1, w2, S2HIMEM #24 406 bne 6f 407 ldr w1, [wp1], #4 408 b 1b 4094: 410 S2LOMEM w2, w2, #8 411 b 8f 412 4135: 414#ifdef __ARMEB__ 415 /* The syndrome value may contain false ones if the string ends 416 * with the bytes 0x01 0x00 417 */ 418 tst w1, #0xff000000 419 itt ne 420 tstne w1, #0x00ff0000 421 tstne w1, #0x0000ff00 422 beq 7f 423#else 424 bics r3, r3, #0xff000000 425 bne 7f 426#endif 427 ldrb w2, [wp2] 428 S2LOMEM t1, w1, #24 429#ifdef __ARMEB__ 430 lsl w2, w2, #24 431#endif 432 b 8f 433 4346: 435 S2LOMEM t1, w1, #24 436 and w2, w2, #LSB 437 b 8f 438 439 /* Critical inner Loop: Block with 2 bytes initial overlap */ 440 .p2align 2 4412: 442 S2HIMEM t1, w1, #16 443 sub r3, w1, b1 444 S2LOMEM t1, t1, #16 445 bic r3, r3, w1 446 cmp t1, w2, S2LOMEM #16 447 bne 4f 448 ands r3, r3, b1, lsl #7 449 it eq 450 ldreq w2, [wp2], #4 451 bne 5f 452 eor t1, t1, w1 453 cmp t1, w2, S2HIMEM #16 454 bne 6f 455 ldr w1, [wp1], #4 456 b 2b 457 4585: 459#ifdef __ARMEB__ 460 /* The syndrome value may contain false ones if the string ends 461 * with the bytes 0x01 0x00 462 */ 463 tst w1, #0xff000000 464 it ne 465 tstne w1, #0x00ff0000 466 beq 7f 467#else 468 lsls r3, r3, #16 469 bne 7f 470#endif 471 ldrh w2, [wp2] 472 S2LOMEM t1, w1, #16 473#ifdef __ARMEB__ 474 lsl w2, w2, #16 475#endif 476 b 8f 477 4786: 479 S2HIMEM w2, w2, #16 480 S2LOMEM t1, w1, #16 4814: 482 S2LOMEM w2, w2, #16 483 b 8f 484 485 /* Critical inner Loop: Block with 1 byte initial overlap */ 486 .p2align 2 4873: 488 and t1, w1, #LSB 489 cmp t1, w2, S2LOMEM #24 490 sub r3, w1, b1 491 bic r3, r3, w1 492 bne 4f 493 ands r3, r3, b1, lsl #7 494 it eq 495 ldreq w2, [wp2], #4 496 bne 5f 497 eor t1, t1, w1 498 cmp t1, w2, S2HIMEM #8 499 bne 6f 500 ldr w1, [wp1], #4 501 b 3b 5024: 503 S2LOMEM w2, w2, #24 504 b 8f 5055: 506 /* The syndrome value may contain false ones if the string ends 507 * with the bytes 0x01 0x00 508 */ 509 tst w1, #LSB 510 beq 7f 511 ldr w2, [wp2], #4 5126: 513 S2LOMEM t1, w1, #8 514 bic w2, w2, #MSB 515 b 8f 5167: 517 mov r0, #0 518 519 /* Restore registers and stack. */ 520 ldrd r6, r7, [sp] 521 ldrd r4, r5, [sp, #8] 522 adds sp, sp, #16 523 .cfi_def_cfa_offset 0 524 .cfi_restore r4 525 .cfi_restore r5 526 .cfi_restore r6 527 .cfi_restore r7 528 529 bx lr 530 5318: 532 and r2, t1, #LSB 533 and r0, w2, #LSB 534 cmp r0, #1 535 it cs 536 cmpcs r0, r2 537 itt eq 538 S2LOMEMEQ t1, t1, #8 539 S2LOMEMEQ w2, w2, #8 540 beq 8b 541 sub r0, r2, r0 542 543 /* Restore registers and stack. */ 544 ldrd r6, r7, [sp] 545 ldrd r4, r5, [sp, #8] 546 adds sp, sp, #16 547 548 bx lr 549END(strcmp) 550